On 8/12/21 10:32 AM, Christoph Hellwig wrote:
> [adding Thomas for a cpu hotplug questions]
>
>> +static void bio_alloc_cache_destroy(struct bio_set *bs)
>> +{
>> + int cpu;
>> +
>> + if (!bs->cache)
>> + return;
>> +
>> + preempt_disable();
>> + cpuhp_state_remove_instance_nocalls(CPUHP_BIO_DEAD, &bs->cpuhp_dead);
>> + for_each_possible_cpu(cpu) {
>> + struct bio_alloc_cache *cache;
>> +
>> + cache = per_cpu_ptr(bs->cache, cpu);
>> + bio_alloc_cache_prune(cache, -1U);
>> + }
>> + preempt_enable();
>
> If I understand the cpu hotplug state machine we should not get any new
> cpu down callbacks after cpuhp_state_remove_instance_nocalls returned,
> so what do we need the preempt disable here for?
I don't think we strictly need it. I can kill it.
>> + /*
>> + * Hot un-plug notifier for the per-cpu cache, if used
>> + */
>> + struct hlist_node cpuhp_dead;
>
> Nit, even if we don't need the cpu up notifaction the node actually
> provides both. So I'd reword the comment drop the _dead from the
> member name.
Right, but we only sign up for the down call.
--
Jens Axboe
